ATAC-seq signal processing and recurrent neural networks can identify RNA polymerase activity

Abstract

Nascent transcription assays are the current gold standard for identifying regions of active transcription, including markers for functional transcription factor (TF) binding. Here we present a signal processingbased model to determine regions of active transcription genome-wide using the simpler assay for transposase-accessible chromatin, followed by high-throughput sequencing (ATAC-seq). The focus of this study is twofold: First, we perform a frequency space analysis of the "signal" generated from ATAC-seq experiments' short reads, at a single-nucleotide resolution, using a discrete wavelet transform. Second, we explore different uses of neural networks to combine this signal with its underlying genome sequence in order to classify ATAC-seq peaks on the presence or absence of bidirectional transcription. We analyze the performance of different data encoding schemes and machine learning architectures, and show how a hybrid signal/sequence representation classified using recurrent neural networks (RNNs) yields the best performance across different cell types.

Combining signal and sequence to detect RNA polymerase initiation in ATAC-seq data

“…The choice of a hybrid encoding scheme for each OCR fixed window resulted from a previous study [27], where we evaluated the performance of many different data encoding schemes and machine learning classifiers. OCRs were evaluated using only the signal at each nucleotide, only the underlying sequence, or a combination of both.…”

“…We combined our sequence embedding and signal into a single vector representation (Fig 1a) that is utilized as input to the GRUs (Fig 1b). A learning rate of 0.0001, a dropout date of 0.1, an embedding layer size of 50 and a hidden layer size of 100 were selected after hyperparameter optimization, from a grid of embedding dimensions [15, 50, In a previous study [27] we evaluated a variety of classifiers and encodings for our RNN and found that the RNN outperformed all other methods at predicting histone marks associated with OCRs that are related to active transcription. For completeness, we briefly summarize the earlier study.…”